multihead_matmul_op.cu 16.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include <paddle/fluid/platform/device_context.h>
16

17
#include <algorithm>
W
Wilber 已提交
18
#include <type_traits>
19

20 21
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/malloc.h"
22
#include "paddle/fluid/operators/math/bert_encoder_functor.h"
W
Wilber 已提交
23
#include "paddle/fluid/platform/float16.h"
24
#include "paddle/phi/core/tensor_utils.h"
25
#include "paddle/phi/kernels/funcs/blas/blas.h"
26 27 28 29 30

namespace paddle {
namespace operators {

template <typename T>
31 32 33 34 35
__global__ void transpose(T *src,
                          T *dst,
                          const int batch_size,
                          const int seq_len,
                          const int head_num,
36 37 38 39 40 41 42 43 44
                          const int size_per_head) {
  int batch_id = blockIdx.x / (head_num * seq_len);
  int seq_id = blockIdx.x % seq_len;
  int head_id = (blockIdx.x % (head_num * seq_len)) / seq_len;
  dst[batch_id * (head_num * seq_len * size_per_head) +
      seq_id * head_num * size_per_head + head_id * size_per_head +
      threadIdx.x] = src[blockIdx.x * size_per_head + threadIdx.x];
}

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
template <typename T>
inline __device__ T add_func(T a, T b);

template <>
__device__ float add_func<float>(float a, float b) {
  return a + b;
}

template <>
__device__ float2 add_func<float2>(float2 a, float2 b) {
  float2 c;
  c.x = a.x + b.x;
  c.y = a.y + b.y;
  return c;
}

template <>
__device__ float4 add_func<float4>(float4 a, float4 b) {
  float4 c;
  c.x = a.x + b.x;
  c.y = a.y + b.y;
  c.z = a.z + b.z;
  c.w = a.w + b.w;
  return c;
69
}
W
Wilber 已提交
70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
#if defined(PADDLE_WITH_CUDA)
template <>
__device__ half2 add_func<half2>(half2 a, half2 b) {
#if __CUDA_ARCH__ >= 530
  return __hadd2(a, b);
#else
  return half2(__float2half(__half2float(a.x) + __half2float(b.x)),
               __float2half(__half2float(b.x) + __half2float(b.y)));
#endif
}

template <>
__device__ half add_func<half>(half a, half b) {
#if __CUDA_ARCH__ >= 530
  return __hadd(a, b);
#else
  return __float2half(__half2float(a) + __half2float(b));
#endif
}
#endif
90 91

template <typename T>
92 93 94
__global__ void TransposeQkvKernel(const int H,
                                   const T *input,
                                   const T *bias,
95
                                   T *output) {
96
  // Input: BxSx3xNxH
W
Wilber 已提交
97
  // Bias: 3xNxH
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
  // Output: 3xBxNxSxH
  int n = threadIdx.y;
  int s = blockIdx.x;
  int b = blockIdx.y;
  int m = blockIdx.z;

  const int N = blockDim.y;
  const int S = gridDim.x;
  const int B = gridDim.y;

  const int NH = N * H;
  const int NHS = NH * S;
  const int in_offset = n * H + m * NH + s * 3 * NH + b * NHS * 3;
  const int bias_offset = m * NH + n * H;
  const int out_offset = s * H + n * S * H + b * NHS + m * NHS * B;

  const int i = threadIdx.x;
  output[out_offset + i] =
      add_func(input[in_offset + i], bias[bias_offset + i]);
}
118

W
Wilber 已提交
119 120 121 122 123 124 125 126 127 128 129
template <typename T>
void TransQKVWithBias(const int batch,
                      const int seq_len,
                      const int head_size,
                      const int head_num,
                      const T *input,
                      const T *bias,
                      T *output,
                      gpuStream_t stream);

template <>
130 131 132 133 134 135 136 137
void TransQKVWithBias(const int batch,
                      const int seq_len,
                      const int head_size,
                      const int head_num,
                      const float *input,
                      const float *bias,
                      float *output,
                      gpuStream_t stream) {
138
  // BxSx3xNxH + 3xNxH -> 3xBxNxSxH
Z
Zhaolong Xing 已提交
139
  int scratch_size = batch * head_num * seq_len * seq_len;
140
  const dim3 grid(seq_len, batch, 3);
Z
Zhaolong Xing 已提交
141 142
  // scratch % 4 == 0 to ensure the alignment
  if (head_size % 4 == 0 && scratch_size % 4 == 0) {
143 144 145 146 147 148 149
    const int h = head_size / 4;
    const float4 *input4 = reinterpret_cast<const float4 *>(input);
    const float4 *bias4 = reinterpret_cast<const float4 *>(bias);
    float4 *output4 = reinterpret_cast<float4 *>(output);
    const dim3 block(h, head_num, 1);

    // limit h * head_num to max block size(1024).
150 151
    PADDLE_ENFORCE_LE(h * head_num,
                      1024,
152 153
                      platform::errors::InvalidArgument(
                          "head_num (%d) * head_size (%d) should <= %d",
154 155 156
                          head_num,
                          head_size,
                          1024 * 4));
157 158
    TransposeQkvKernel<float4>
        <<<grid, block, 0, stream>>>(h, input4, bias4, output4);
Z
Zhaolong Xing 已提交
159
  } else if (head_size % 2 == 0 && scratch_size % 2 == 0) {
160 161 162 163 164 165
    const int h = head_size / 2;
    const float2 *input2 = reinterpret_cast<const float2 *>(input);
    const float2 *bias2 = reinterpret_cast<const float2 *>(bias);
    float2 *output2 = reinterpret_cast<float2 *>(output);
    const dim3 block(h, head_num, 1);
    // limit h * head_num to max block size(1024).
166 167
    PADDLE_ENFORCE_LE(h * head_num,
                      1024,
168 169
                      platform::errors::InvalidArgument(
                          "head_num (%d) * head_size (%d) should <= %d",
170 171 172
                          head_num,
                          head_size,
                          1024 * 2));
173 174
    TransposeQkvKernel<float2>
        <<<grid, block, 0, stream>>>(h, input2, bias2, output2);
175 176 177
  } else {
    const dim3 block(head_size, head_num, 1);
    // limit head_size * head_num to max block size(1024).
178 179
    PADDLE_ENFORCE_LE(head_size * head_num,
                      1024,
180 181
                      platform::errors::InvalidArgument(
                          "head_num (%d) * head_size (%d) should <= %d",
182 183 184
                          head_num,
                          head_size,
                          1024));
185 186
    TransposeQkvKernel<float>
        <<<grid, block, 0, stream>>>(head_size, input, bias, output);
187 188
  }
}
189

W
Wilber 已提交
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238
#if defined(PADDLE_WITH_CUDA)
template <>
void TransQKVWithBias(const int batch,
                      const int seq_len,
                      const int head_size,
                      const int head_num,
                      const platform::float16 *input,
                      const platform::float16 *bias,
                      platform::float16 *output,
                      gpuStream_t stream) {
  // BxSx3xNxH + 3xNxH -> 3xBxNxSxH
  int scratch_size = batch * head_num * seq_len * seq_len;
  const dim3 grid(seq_len, batch, 3);
  if (head_size % 2 == 0 && scratch_size % 2 == 0) {
    const int h = head_size / 2;
    const half2 *input2 = reinterpret_cast<const half2 *>(input);
    const half2 *bias2 = reinterpret_cast<const half2 *>(bias);
    half2 *output2 = reinterpret_cast<half2 *>(output);
    const dim3 block(h, head_num, 1);
    // limit h * head_num to max block size(1024).
    PADDLE_ENFORCE_LE(h * head_num,
                      1024,
                      platform::errors::InvalidArgument(
                          "head_num (%d) * head_size (%d) should <= %d",
                          head_num,
                          head_size,
                          1024 * 2));
    TransposeQkvKernel<half2>
        <<<grid, block, 0, stream>>>(h, input2, bias2, output2);
  } else {
    const dim3 block(head_size, head_num, 1);
    const half *input_half = reinterpret_cast<const half *>(input);
    const half *bias_half = reinterpret_cast<const half *>(bias);
    half *output_half = reinterpret_cast<half *>(output);

    // limit head_size * head_num to max block size(1024).
    PADDLE_ENFORCE_LE(head_size * head_num,
                      1024,
                      platform::errors::InvalidArgument(
                          "head_num (%d) * head_size (%d) should <= %d",
                          head_num,
                          head_size,
                          1024));
    TransposeQkvKernel<half><<<grid, block, 0, stream>>>(
        head_size, input_half, bias_half, output_half);
  }
}
#endif

F
feng_shuai 已提交
239 240
inline int round_up(int seq_len, int multiple = 32) {
  PADDLE_ENFORCE_GT(
241 242
      multiple,
      0,
F
feng_shuai 已提交
243 244 245 246 247 248
      platform::errors::InvalidArgument(
          "multiple should be a positive number,but it's (%d)", multiple));
  return ((seq_len + multiple - 1) / multiple) * multiple;
}

template <typename T>
249 250 251
__global__ void broadcast(const T *src,
                          T *dst,
                          const int seq_len,
F
feng_shuai 已提交
252 253 254 255 256 257 258 259
                          const int head_num) {
  int batch_id = blockIdx.x / (head_num * seq_len);
  int dst_offset = blockIdx.x * seq_len;
  if (threadIdx.x < seq_len) {
    dst[threadIdx.x + dst_offset] = src[threadIdx.x + batch_id * seq_len];
  }
}

260 261 262 263 264 265 266 267 268 269 270 271 272
template <typename T>
__global__ void broadcast_batch_head_number(const T *src,
                                            T *dst,
                                            const int batch_size,
                                            const int seq_len,
                                            const int head_num) {
  int src_seq_id = blockIdx.x % seq_len;
  int dst_offset = blockIdx.x * seq_len;
  if (threadIdx.x < seq_len) {
    dst[threadIdx.x + dst_offset] = src[threadIdx.x + src_seq_id * seq_len];
  }
}

273
template <typename DeviceContext, typename T>
274
class MultiHeadMatMulV2Kernel : public framework::OpKernel<T> {
275 276
 public:
  void Compute(const framework::ExecutionContext &context) const override {
277 278 279 280
    auto *input = context.Input<phi::DenseTensor>("Input");
    auto *w = context.Input<phi::DenseTensor>("W");
    auto *bias = context.Input<phi::DenseTensor>("Bias");
    auto *bias_qk = context.Input<phi::DenseTensor>("BiasQK");
281

282 283 284
    auto *input_d = input->data<T>();
    auto *w_d = w->data<T>();
    auto *bias_d = bias->data<T>();
F
feng_shuai 已提交
285
    auto *bias_qk_d = bias_qk ? bias_qk->data<T>() : nullptr;
286 287 288 289 290
    T scale = static_cast<T>(context.Attr<float>("alpha"));

    int head_number = context.Attr<int>("head_number");
    // compute q*k with eltadd
    auto &device_ctx = context.template device_context<DeviceContext>();
F
feng_shuai 已提交
291
    auto stream = device_ctx.stream();
292 293 294 295 296 297 298
    // should be (B * S * hidden)
    auto input_dims = input->dims();
    // shouble be (hidden * 3 * all_head_size)
    auto w_dims = w->dims();
    int batch = input_dims[0];
    int seq_len = input_dims[1];
    int hidden = input_dims[2];
299
    phi::DenseTensor temp_bias_tensor;
F
feng_shuai 已提交
300
    // if bias_qk is[batch, 1, 1, seq_len], the bias_qk_d need to be broadcasted
F
feng_shuai 已提交
301
    if (bias_qk && bias_qk->numel() == (batch * seq_len)) {
302
      VLOG(4) << "Do broadcasted bias_qk from [batch, 1, 1, seq_len]";
F
feng_shuai 已提交
303
      temp_bias_tensor.Resize({batch * head_number * seq_len * seq_len});
W
Wilber 已提交
304 305
      auto *temp_qk_bias = device_ctx.template Alloc<T>(
          &temp_bias_tensor, temp_bias_tensor.numel() * sizeof(T));
F
feng_shuai 已提交
306 307
      int grid = batch * head_number * seq_len;
      int block = round_up(seq_len);
308 309
      broadcast<<<grid, block, 0, stream>>>(
          bias_qk_d, temp_qk_bias, seq_len, head_number);
F
feng_shuai 已提交
310 311
      bias_qk_d = static_cast<const T *>(temp_qk_bias);
    }
312 313 314 315 316 317 318 319 320 321 322 323 324
    // if bias_qk is[1, 1, seq_len, seq_len], the bias_qk_d need to be
    // broadcasted
    if (bias_qk && bias_qk->numel() == (1 * seq_len * seq_len)) {
      VLOG(4) << "do broadcasted bias_qk from  [1, 1, seq_len, seq_len]";
      temp_bias_tensor.Resize({batch * head_number * seq_len * seq_len});
      auto *temp_qk_bias = device_ctx.template Alloc<T>(
          &temp_bias_tensor, temp_bias_tensor.numel() * sizeof(T));
      int grid = batch * head_number * seq_len;
      int block = round_up(seq_len);
      broadcast_batch_head_number<<<grid, block, 0, stream>>>(
          bias_qk_d, temp_qk_bias, batch, seq_len, head_number);
      bias_qk_d = static_cast<const T *>(temp_qk_bias);
    }
F
feng_shuai 已提交
325 326 327
    if (!bias_qk) {
      int size = batch * head_number * seq_len * seq_len;
      temp_bias_tensor.Resize({size});
W
Wilber 已提交
328 329
      auto *temp_qk_bias = device_ctx.template Alloc<T>(
          &temp_bias_tensor, temp_bias_tensor.numel() * sizeof(T));
F
feng_shuai 已提交
330 331 332 333 334 335 336
#ifdef PADDLE_WITH_HIP
      hipMemset(temp_qk_bias, 0, sizeof(float) * size);
#else
      cudaMemset(temp_qk_bias, 0, sizeof(float) * size);
#endif
      bias_qk_d = static_cast<const T *>(temp_qk_bias);
    }
337 338 339
    int all_head_size = w_dims[2];
    int head_size = all_head_size / head_number;

340
    auto *out = context.Output<phi::DenseTensor>("Out");
341
    out->Resize({batch, seq_len, all_head_size});
W
Wilber 已提交
342 343
    auto *output_d =
        device_ctx.template Alloc<T>(out, out->numel() * sizeof(T));
344

345
    // (B*S, hidden)
346
    const phi::DenseTensor input_matrix =
347
        phi::ReshapeToMatrix(*input, 2 /*x_num_col_dims */);
348
    // (hidden, 3 * all_head_size)
349
    const phi::DenseTensor w_matrix =
350
        phi::ReshapeToMatrix(*w, 1 /*y_num_col_dims*/);
351

352
    phi::DenseTensor temp_out_tensor;
353
    auto temp_out_dims =
354
        phi::make_ddim({batch, seq_len, 3, head_number, head_size});
355
    temp_out_tensor.Resize(
356
        {batch * seq_len, phi::product(temp_out_dims) / (batch * seq_len)});
W
Wilber 已提交
357 358
    auto *temp_out_data = device_ctx.template Alloc<T>(
        &temp_out_tensor, temp_out_tensor.numel() * sizeof(T));
359 360

    // (B * S, hidden) * (hidden, 3 * N * H) -> (B * S * 3 * N * H)
L
Leo Chen 已提交
361
    auto blas = phi::funcs::GetBlas<phi::GPUContext, T>(device_ctx);
362
    blas.MatMul(input_matrix, w_matrix, &temp_out_tensor);
363 364
    VLOG(2) << "(B * S, hidden) * (hidden, 3 * N * H) -> (B * S * 3 * N * H)";
    VLOG(2) << temp_out_tensor;
365 366
    // temp_out_tensor.Resize(temp_out_dims);

367
    phi::DenseTensor multihead_temp_tensor;
368 369 370
    // B * head_number * S * S * 1 + B * S * 3 * N * H
    int scratch_size = batch * head_number * seq_len * seq_len * 1;
    multihead_temp_tensor.Resize({scratch_size + temp_out_tensor.numel()});
W
Wilber 已提交
371 372 373
    auto *multihead_temp_data = device_ctx.template Alloc<T>(
        &multihead_temp_tensor, multihead_temp_tensor.numel() * sizeof(T));

374 375 376 377 378
    auto *qkptr = multihead_temp_data;
    auto *tptr = multihead_temp_data + scratch_size;

    // Do the transpose with bias.
    // BxSx3xNxH => tptr: 3xBxNxSxH.
379 380 381 382 383 384 385 386
    TransQKVWithBias(batch,
                     seq_len,
                     head_size,
                     head_number,
                     temp_out_data,
                     bias_d,
                     tptr,
                     stream);
W
Wilber 已提交
387 388 389 390 391 392 393 394 395
    if (std::is_same<T, platform::float16>::value) {
      math::MultiHeadGPUComputeFunctor<half> multihead_compute_func;
      multihead_compute_func(device_ctx,
                             batch,
                             seq_len,
                             head_number,
                             head_size,
                             reinterpret_cast<half *>(qkptr),
                             reinterpret_cast<const half *>(bias_qk_d),
F
feng_shuai 已提交
396
                             false,
W
Wilber 已提交
397 398 399 400 401 402 403 404 405 406 407 408
                             reinterpret_cast<half *>(tptr),
                             __float2half(static_cast<float>(scale)),
                             __float2half(0.0));
    } else {
      math::MultiHeadGPUComputeFunctor<T> multihead_compute_func;
      multihead_compute_func(device_ctx,
                             batch,
                             seq_len,
                             head_number,
                             head_size,
                             qkptr,
                             bias_qk_d,
F
feng_shuai 已提交
409
                             false,
W
Wilber 已提交
410 411 412 413
                             tptr,
                             scale,
                             T(0.0));
    }
414 415 416

    int grid = batch * head_number * seq_len;
    int block = head_size;
417 418
    transpose<T><<<grid, block, 0, stream>>>(
        tptr, output_d, batch, seq_len, head_number, head_size);
419 420 421 422 423 424 425
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
W
Wilber 已提交
426 427 428 429 430 431
#if defined(PADDLE_WITH_CUDA) && CUDA_VERSION >= 10000
REGISTER_OP_CUDA_KERNEL(
    multihead_matmul,
    ops::MultiHeadMatMulV2Kernel<phi::GPUContext, paddle::platform::float16>,
    ops::MultiHeadMatMulV2Kernel<phi::GPUContext, float>);
#else
L
Leo Chen 已提交
432 433
REGISTER_OP_CUDA_KERNEL(multihead_matmul,
                        ops::MultiHeadMatMulV2Kernel<phi::GPUContext, float>);
W
Wilber 已提交
434
#endif